NVIDIA NeMo Framework

Litlhaloso

• Lebitso la Sehlahiswa: NVIDIA NeMo Framework
• Libaka tse Amehileng: Windows, Linux, macOS
• Mefuta e Amehileng: Liphetolelo tsohle pele ho 24
• Tšireletso kotsing: CVE-2025-23360
• Lintlha tsa Motheo tsa Tekolo ea Kotsi: 7.1 (CVSS v3.1)

Litaelo tsa Tšebeliso ea Sehlahisoa

Kemiso ea ntlafatso ea ts'ireletso:
Ho sireletsa sistimi ea hau, latela mehato ena:

1. Khoasolla tokollo ea morao-rao ho tsoa leqepheng la Nemo-Framework-Launcher Releases ho GitHub.
2. Eya ho NVIDIA Product Security bakeng sa lintlha tse ling.

Lintlha tsa ntlafatso ea Ts'ireletso:
Ntlafatso ea ts'ireletso e sebetsana le ho ba kotsing ho NVIDIA NeMo Framework e ka lebisang ts'ebetsong ea khoutu le data t.ampho kheloha.

Ntlafatsa Software:
Haeba u sebelisa tokollo ea pejana ea lekala, ho khothaletsoa ho ntlafatsa tokollo ea morao-rao ea lekala ho rarolla bothata ba ts'ireletso.

Fetileview

NVIDIA NeMo Framework ke moralo oa AI o ka senyehang hape o hlahisoang ke maru o etselitsoeng bafuputsi le bahlahisi ba sebetsang ho bona. Mehlala e Meholo ea Puo, Multimodal, le Puo ea AI (mohl Kamohelo ea Puo e Ikemetseng le Mongolo ho ea ho Puo). E thusa basebelisi ho theha, ho etsa, le ho sebelisa mefuta e mecha ea AI ka ho sebelisa khoutu e teng le libaka tsa tlhahlobo tse koetlisitsoeng esale pele.

Litaelo tsa ho seta: Kenya NeMo Framework

Mehlala e Meholo ea Lipuo le Mehlala ea Multimodal
NeMo Framework e fana ka tshehetso ya ho qetela bakeng sa ho ntshetsapele Mehlala e Kgolo ya Dipuo (LLMs) le Multimodal Models (MMs). E fana ka monyetla oa ho sebelisoa sebakeng sa marang-rang, setsing sa data, kapa le mofani oa hau oa maru eo u mo ratang. E boetse e ts'ehetsa ts'ebetso ho SLURM kapa Kubernetes e lumelletsoeng tikoloho.

Khatiso ea Lintlha
Mohlokomeli oa NeMo [1] ke laebrari ea Python e kenyelletsang lethathamo la li-module tsa merafo ea data le tlhahiso ea data ea maiketsetso. Li na le sekhahla ebile li ntlafalitsoe bakeng sa li-GPU, li li etsa tse loketseng ho boloka lintlha tsa puo ea tlhaho ho koetlisa kapa ho lokisa li-LLM. Ka NeMo Curator, o ka khona ho hula mongolo oa boleng bo holimo ho tsoa ho tse tala web mehloli ea data.

Koetliso le ho Itloaetsa

NeMo Framework e fana ka lisebelisoa tsa koetliso e sebetsang hantle le ho e etsa ka mokhoa o ikhethileng oa LLMs le mefuta ea Multimodal. E kenyelletsa litlhophiso tsa kamehla bakeng sa ho seta sehlopha sa khomphutha, ho khoasolla data, le li-hyperparameter tsa mohlala, tse ka lokisoang ho koetlisa li-dataset tse ncha le mefuta. Ntle le koetliso ea pele, NeMo e ts'ehetsa mekhoa ea Supervised Fine-Tuning (SFT) le Parameter Efficient Fine-Tuning (PEFT) joalo ka LoRA, Ptuning, le tse ling.

Likhetho tse peli li teng ho qala koetliso ho NeMo - ho sebelisa sebopeho sa NeMo 2.0 API kapa ka NeMo Run.

• Ka NeMo Run (E khothalelitsoe): NeMo Run e fana ka sebopeho sa ho tsamaisa tlhophiso, ts'ebetso le taolo ea liteko libakeng tse fapaneng tsa likhomphutha. Sena se kenyelletsa ho qala mesebetsi setsing sa hau sa mosebetsi sebakeng sa heno kapa ho lihlopha tse kholo - ka bobeli SLURM e lumelletsoe kapa Kubernetes sebakeng sa maru.
  • Koetliso ea pele le PEFT Quickstart ka NeMo Run
• Ho sebelisa NeMo 2.0 API: Mokhoa ona o sebetsa hantle ka mokhoa o bonolo oa ho seta o kenyelletsang mehlala e menyenyane, kapa haeba u thahasella ho ngola dataloader ea hau ea tloaelo, li-loops tsa koetliso, kapa ho fetola likarolo tsa mohlala. E u fa monyetla oa ho feto-fetoha le maemo le taolo holim'a litlhophiso, 'me e etsa hore ho be bonolo ho atolosa le ho iketsetsa litlhophiso ka mokhoa oa lenaneo.
  • Training Quickstart ka NeMo 2.0 API
  • Ho falla ho tloha NeMo 1.0 ho ea ho NeMo 2.0 API

Ho tsamaisana

• NeMo-Aligner [1] ke sesebelisoa sa scalable bakeng sa ho tsamaisana hantle le mohlala. Setsi sa lithulusi se na le ts'ehetso bakeng sa li-algorithms tsa boemo ba morao-rao tsa ho lumellana ha mohlala tse kang SteerLM, DPO, Reinforcement Learning from Human Feedback (RLHF), le tse ling tse ngata. Li-algorithms tsena li thusa basebelisi ho hokahanya mefuta ea lipuo hore e bolokehe, e se be kotsi le ho thusa.
• Libaka tsohle tsa tlhahlobo tsa NeMo-Aligner li tsamaisana hantle le NeMo ecosystem, e lumellang hore ho be le tlhophiso e eketsehileng le ho tsamaisoa ha maikutlo.

Ts'ebetso ea mohato ka mohato ea mekhahlelo e meraro ea RLHF mofuteng o monyane oa GPT-2B:

• Koetliso ea SFT
• Koetliso ea mohlala oa moputso
• Koetliso ea PPO

Ntle le moo, re bonts'a ts'ehetso bakeng sa mekhoa e meng e fapaneng ea ho hokahanya:

• DPO: algorithm ea ho tsamaisana ha bobebe ha e bapisoa le RLHF e nang le ts'ebetso e bonolo ea tahlehelo.
• Boiketsi ba ho Bapala Tokiso e Ntle (SPIN)
• SteerLM: mokhoa o thehiloeng ho conditioned-SFT, e nang le tlhahiso e tsitsitseng.

Sheba litokomane bakeng sa lintlha tse ling: Litokomane tsa ho tsamaisana

Mehlala ea Multimodal

• NeMo Framework e fana ka software e ntlafalitsoeng ea ho koetlisa le ho tsamaisa mefuta ea morao-rao ea li-multimodal ho pholletsa le mekhahlelo e 'maloa: Mehlala ea Puo ea Multimodal, Metheo ea Puo ea Pono, Mehlala ea Mongolo ho ea ho Sets'oants'o, le ho feta 2D Generation e sebelisang Neural Radiance Fields (NeRF).
• Sehlopha se seng le se seng se etselitsoe ho khotsofatsa litlhoko le tsoelo-pele e itseng tšimong, ho sebelisa mefuta e metle ea ho sebetsana le mefuta e mengata ea data, ho kenyeletsoa mongolo, litšoantšo le mefuta ea 3D.

Phatlalatso le Tlhahiso
NeMo Framework e fana ka litsela tse fapaneng tsa tlhahiso ea LLM, e sebetsanang le maemo a fapaneng a thomello le litlhoko tsa ts'ebetso.

E tsamaisa ka NVIDIA NIM

• NeMo Framework e hokahana ka mokhoa o ts'oanang le lisebelisoa tsa phepelo ea maemo a khoebo ka NVIDIA NIM. Khokahano ena e tsamaisoa ke NVIDIA TensorRT-LLM, e netefatsa hore ho na le tšibollo e ntlafalitsoeng le e ka senyehang.
• Bakeng sa tlhaiso-leseling e batsi ka NIM, etela NVIDIA websebaka.

Sebelisa ka TensorRT-LLM kapa vLLM

• NeMo Framework e fana ka lingoloa le li-API ho romella mefuta e meng ho lilaebraring tse peli tse ntlafalitsoeng, TensorRT-LLM le vLLM, le ho tsamaisa mofuta o romelloang kantle ho naha ka NVIDIA Triton Inference Server.
• Bakeng sa maemo a hlokang ts'ebetso e ntlafalitsoeng, mefuta ea NeMo e ka sebelisa TensorRT-LLM, laebrari e ikhethileng bakeng sa ho potlakisa le ho ntlafatsa tlhahiso ea LLM ho NVIDIA GPUs. Ts'ebetso ena e kenyelletsa ho fetolela mefuta ea NeMo hore e be sebopeho se lumellanang le TensorRT-LLM ho sebelisa mojule oa nemo.export.
  • Ho tsamaisoa ha LLM ho Felletseview
  • Kenya Mefuta e Meholo ea Nemo ea Puo e nang le NIM
  • Kenya Mefuta e Meholo ea Puo ea NeMo ka TensorRT-LLM
  • Kenya Mefuta e Meholo ea Nemo ea Puo e nang le vLLM

Mehlala e Tšehetsoeng

Mehlala e Meholo ea Puo

Mehlala e Meholo ea Puo

Mehlala e Meholo ea Puo	Ho ikoetlisa pele & SFT	PEFT	Ho tsamaisana	Khokahano ea Koetliso ea FP8	TRT/TRTLLM	Fetolela Ho & Ho Tsoa ho Hugging Sefahleho	Tlhahlobo
Llama3 8B/70B, Llama3.1 405B	Ee	Ee	x	Ee (e netefalitsoe ka mokhoa o itseng)	Ee	Ka bobeli	Ee
Motsoako oa 8x7B/8x22B	Ee	Ee	x	Ee (ha ea netefatsoa)	Ee	Ka bobeli	Ee
Nemotron 3 8B	Ee	x	x	Ee (ha ea netefatsoa)	x	Ka bobeli	Ee
Nemotron 4 340B	Ee	x	x	Ee (ha ea netefatsoa)	x	Ka bobeli	Ee
Baichuan2 7B	Ee	Ee	x	Ee (ha ea netefatsoa)	x	Ka bobeli	Ee
ChatGLM3 6B	Ee	Ee	x	Ee (ha ea netefatsoa)	x	Ka bobeli	Ee
Gemma 2B/7B	Ee	Ee	x	Ee (ha ea netefatsoa)	Ee	Ka bobeli	Ee
Gemma2 2B/9B/27B	Ee	Ee	x	Ee (ha ea netefatsoa)	x	Ka bobeli	Ee
Mamba2 130M/370M/780M/1.3B/2.7B/8B/ Hybrid-8B	Ee	Ee	x	Ee (ha ea netefatsoa)	x	x	Ee
Phi3 mini 4k	x	Ee	x	Ee (ha ea netefatsoa)	x	x	x
Qwen2 0.5B/1.5B/7B/72B	Ee	Ee	x	Ee (ha ea netefatsoa)	Ee	Ka bobeli	Ee
StarCoder 15B	Ee	Ee	x	Ee (ha ea netefatsoa)	Ee	Ka bobeli	Ee
StarCoder2 3B/7B/15B	Ee	Ee	x	Ee (ha ea netefatsoa)	Ee	Ka bobeli	Ee
BERT 110M/340M	Ee	Ee	x	Ee (ha ea netefatsoa)	x	Ka bobeli	x
T5 220M/3B/11B	Ee	Ee	x	x	x	x	x

 

Pono Mehlala ea Puo

 

Embedding Models

 

Mehlala ea Motheo oa Lefatše

Hlokomela
NeMo e boetse e ts'ehetsa ho ikoetlisa pele bakeng sa meaho e fapaneng le ea autoregressive text2world mehlala ea motheo.

Puo ea AI

Ho nts'etsapele mefuta ea puisano ea AI ke ts'ebetso e rarahaneng e kenyelletsang ho hlalosa, ho aha, le ho koetlisa mehlala ka har'a libaka tse itseng. Ts'ebetso ena hangata e hloka liphetolelo tse 'maloa ho fihlela boemo bo holimo ba ho nepahala. Hangata e kenyelletsa ho pheta-pheta makhetlo a mangata ho fihlela ho nepahala ho phahameng, ho hlophisa hantle mesebetsing e fapaneng le data e ikhethileng ea domain, ho netefatsa ts'ebetso ea koetliso, le ho lokisa mehlala ea ho tsamaisoa ha maikutlo.

NeMo Framework e fana ka ts'ehetso bakeng sa koetliso le tloaelo ea mefuta ea Speech AI. Sena se kenyelletsa mesebetsi e kang Automatic Speech Recognition (ASR) le Text-to-Speech (TTS) synthesis. E fana ka phetoho e bonolo ho isa tlhahiso ea boemo ba khoebo le NVIDIA Riva. Ho thusa bahlahisi le bafuputsi, NeMo Framework e kenyelletsa libaka tsa tlhahlobo tsa morao-rao tse koetlisitsoeng esale pele, lisebelisoa tsa ts'ebetso ea data ea puo e ka hlahisoang hape, le likarolo tsa tlhahlobo e kopanetsoeng le tlhahlobo ea li-database tsa puo. Likarolo tsa NeMo Framework for Speech AI ke tse latelang:

Koetliso le ho Itloaetsa
NeMo Framework e na le tsohle tse hlokahalang ho koetlisa le ho etsa mefuta ea lipuo (ASR, Karolelano ea Lipuo, Tlhokomeliso ea Sebui, Puisano ea libui, le TTS) ka tsela e ka ikatisang.

Mehlala e koetlisitsoeng esale pele ea SOTA

• NeMo Framework e fana ka litlolo tsa morao-rao le libaka tsa tlhahlobo tse koetlisitsoeng esale pele tsa tse 'maloa ASR le TTS mehlala, hammoho le litaelo tsa mokhoa oa ho li kenya.
• Lisebelisoa tsa Puo
• NeMo Framework e fana ka lisebelisoa tse sebetsang bakeng sa ho theha mefuta ea ASR le TTS, ho kenyelletsa:
  • NeMo Forced Aligner (NFA) bakeng sa ho hlahisa nako ea token-, word- and segment-level timesamps ea puo ka molumo o sebelisang mefuta ea Nemo's CTC-based Automatic Speech Recognition.
  • Speech Data Processor (SDP), sesebelisoa sa ho nolofatsa ts'ebetso ea data ea puo. E u lumella ho emela ts'ebetso ea ts'ebetso ea data ho config file, ho fokotsa khoutu ea boilerplate le ho lumella ho ikatisa le ho arolelana.
  • Speech Data Explorer (SDE), e thehiloeng ho Dash web ts'ebeliso ea tlhahlobo e kopanetsoeng le tlhahlobo ea pokello ea lintlha tsa puo.
  • Sesebelisoa sa ho etsa li-dataset e fanang ka tshebetso ya ho ikamahanya le modumo o molelele files le lingoliloeng tse tsamaellanang le ho li arola likotoana tse khutšoane tse loketseng koetliso ea mohlala ea Automatic Speech Recognition (ASR).
  • Papiso Tool bakeng sa Mehlala ea ASR ho bapisa likhakanyo tsa mefuta e fapaneng ea ASR ka nepo ea mantsoe le boemo ba polelo.
  • Mohlahlobi oa ASR bakeng sa ho lekola ts'ebetso ea mefuta ea ASR le likarolo tse ling tse joalo ka Pono ea Ketsahalo ea Lentsoe.
  • Sengoloa Normalization Tool bakeng sa ho fetolela mongolo ho tloha mofuteng o ngotsweng ho ya ho o buuwang le ka tsela e fapaneng (mohlala, “31st” vs “mashome a mararo pele”).
• Tsela ea ho Tsamaisa
• Mefuta ea NeMo e koetlisitsoeng kapa e etselitsoeng mokhoa oa ho sebelisa NeMo Framework e ka ntlafatsoa le ho sebelisoa le NVIDIA Riva. Riva e fana ka lijana le lichate tsa Helm tse etselitsoeng ka ho khetheha ho iketsetsa mehato ea ho kenya konopo ea push-button.

Lisebelisoa tse ling

GitHub Repos

• NeMo: Sebaka sa polokelo ea mantlha ea NeMo Framework
• NeMo–Matha: Sesebelisoa sa ho hlophisa, ho qala le ho laola liteko tsa hau tsa ho ithuta mochini.
• NeMo-Aligner: Scalable toolkit bakeng sa ho tsamaisana hantle le mohlala
• NeMo-Curator: Scalable data pre-processing le curation toolkit bakeng sa LLMs

Ho Fumana Thuso
Kopana le sechaba sa NeMo, botsa lipotso, fumana tšehetso, kapa tlaleha liphoso.

• Nemo Lipuisano
• Litaba tsa NeMo

Lipuo tsa Mananeo le Mekhahlelo

• Python: Sebopeho se seholo sa ho sebelisa NeMo Framework
• Pytorch: NeMo Framework e hahiloe holim'a PyTorch

Lilaesense

• NeMo Github repo e na le laesense tlasa laesense ea Apache 2.0
• NeMo Framework e na le lengolo la tumello tlas'a NVIDIA AI PRODUCT AGREEMENT. Ka ho hula le ho sebelisa setshelo, o amohela lipehelo le lipehelo tsa laesense ena.
• Sets'oants'o sa NeMo Framework se na le lisebelisoa tsa Llama tse laoloang ke Tumellano ea License ea Sechaba ea Meta Llama3.

Mengolo e botlaaseng ba leqephe
Hajoale, tšehetso ea NeMo Curator le NeMo Aligner bakeng sa mefuta ea Multimodal ke mosebetsi o ntseng o tsoela pele 'me o tla fumaneha haufinyane.

LBH

P: Nka hlahloba joang hore na tsamaiso ea ka e angoa ke ho ba kotsing?
A: O ka hlahloba hore na sistimi ea hau e angoa ke ho netefatsa mofuta oa NVIDIA NeMo Framework e kentsoeng. Haeba e ka tlase ho mofuta oa 24, sistimi ea hau e kanna ea ba kotsing.

P: Ke mang ea tlalehileng taba ea ts'ireletso CVE-2025-23360?
A: Taba ea ts'ireletso e tlalehiloe ke Or Peles - JFrog Security. NVIDIA e ananela tlatsetso ea bona.

P: Nka fumana litemoso tsa litaba tsa ts'ireletso joang nakong e tlang?
A: Etela leqephe la Tšireletso ea Sehlahisoa sa NVIDIA ho ingolisa ho litsebiso tsa litaba tsa ts'ireletso le ho lula u tseba ka lintlafatso tsa ts'ireletso ea sehlahisoa.

Litokomane / Lisebelisoa

NVIDIA NeMo Framework [pdf] Bukana ea Mosebelisi NeMo Framework, NeMo, Framework

Litšupiso

• Bukana ea Mosebelisi